Abstract
Brain-machine interfaces (BMIs) have been applied as new rehabilitation tools for motor disabled individuals. Active involvement of cerebral activity has been shown to enhance neuroplasticity and thus to restore mobility. Various studies have focused on the detection of upper-limb movement intention, while fewer studies have investigated lower-limb movement intention decoding. This study presents a BMI to decode self-paced lower-limb movement intention, with 10 healthy subjects participating in the experiment. We varied four influence factors including the movement type (dorsiflexion or plantar flexion), the limb side (left or right leg), the processing method (time-series analysis based on MRCP or frequency-domain estimation based on SMR), and the frequency band to estimate both single-trial and sample-based performance. The average detection latency was -0.334 ± 0.216 s. An average area under the curve (AUC) of 91.0 ± 3.5% and 68.2 ± 4.6% was obtained for the MRCP-based and SMR-based methods, respectively. The best performance was yielded from plantar flexion with the left leg using time-series analysis on the MRCP band.
Introduction
Brain-Machine Interfaces (BMIs) are transformative tools for neurorehabilitation, particularly for patients recovering from stroke or living with paralysis. To promote neuroplasticity—the brain's ability to rewire itself—it is crucial that a robotic gait trainer assists a patient exactly when they intend to move. While upper-limb intention (reaching, grasping) has been studied extensively, lower-limb decoding remains a challenge. Most existing studies focus on a single type of movement or a specific frequency band. This paper addresses the lack of a comprehensive "factor analysis," investigating how movement type (flexing the foot up vs. down), the specific leg used, and different signal processing frameworks affect the accuracy and timing of a BMI "switch."
Methods and Experimental Protocol
The researchers recruited 10 healthy, right-handed subjects (average age ~25). Each subject performed approximately 300 trials of self-paced ankle movements (dorsiflexion and plantar flexion) across both legs. A specialized robotic device called the legoPress was used to monitor leg movements. 32-channel EEG (brain activity) and EMG (muscle activity) were recorded to establish the ground truth for when a movement actually started.
The core of the approach was a head-to-head comparison of two primary EEG features:
- Movement-Related Cortical Potentials (MRCPs): These are slow, time-domain voltage shifts that occur as the brain prepares to move.
- Sensory Motor Rhythms (SMRs): These are frequency-domain changes (like the "mu" or "beta" rhythms) where the power of specific oscillations drops during movement preparation.
Canonical Variant Analysis (CVA) was utilized to extract and rank the most discriminative features, and Linear Discriminant Analysis (LDA) was used to classify the brain's state as either "idle" or "preparing to move."
Key Results and Insights
The study revealed several critical findings regarding the optimal design of a lower-limb BMI "brain switch":
- Superiority of Time-Domain Analysis: The most significant finding was that MRCP-based methods (time-domain) vastly outperformed SMR-based methods (frequency-domain) for detecting the onset of movement. The MRCP-based method achieved an average Area Under the Curve (AUC) of 91.0%, compared to only 68.2% for the SMR-based method.
- Consistency vs. Specificity: MRCP features were remarkably consistent across different subjects, localized primarily in the central motor cortex (the Cz electrode). In contrast, SMR features were highly subject-specific, making them harder to generalize.
- The "Sweet Spot" for Timing: The system detected intention at an average of -0.334 seconds before muscle activity. This lead time is critical; it provides just enough window for a robotic exoskeleton to overcome mechanical latency and move in sync with the human user.
- Asymmetry in Performance: Interestingly, the best performance was yielded from plantar flexion (pointing the toes) with the left leg.
Discussion and Future Directions
While promising, several hurdles remain before clinical application. The study was conducted on healthy subjects; patients with stroke or ALS may have significantly altered or weakened EEG signals, which might require more robust adaptive algorithms. Furthermore, this "offline" analysis must be transitioned to "online" real-time systems, which introduces challenges with causal filtering. Future work will focus on integrating this "switch" into closed-loop systems where the user receives immediate feedback from the robot to enhance the learning effect.
Summary updated